Public administrative advisory services, such as the Basic Pension Advisory Service, require accurate policy interpretation and step-by-step information collection based on individual user conditions. However, rule-based chatbots struggle to handle di...
Public administrative advisory services, such as the Basic Pension Advisory Service, require accurate policy interpretation and step-by-step information collection based on individual user conditions. However, rule-based chatbots struggle to handle diverse user situations due to their reliance on predefined scenarios, while general-purpose large language model-based chatbots lack explicit mechanisms for policy reasoning and dialogue flow control, limiting their applicability to policy-sensitive administrative counseling tasks.
To address these limitations, this study proposes a LLM-based goal-oriented dialogue system for Basic Pension Advisory Service, built on an NLU–Planner–NLG architecture. In the proposed system, user utterances are structured into policy-relevant slots, and a LLM-based Planner dynamically determines whether to request additional information or generate a final response by jointly considering the user utterance, the current dialogue state, and an external policy knowledge base constructed from official policy documents and official Q&A documents. The system was evaluated using 50 scenario-based Basic Pension advisory dialogues and compared with a general GPT baseline model and a persona-conditioned GPT baseline model. Experimental results show that the proposed model achieves superior performance in accuracy and conciseness, demonstrating its effectiveness in managing policy-driven dialogue flows and supporting reliable Basic Pension Advisory Service.